Blocking Designs - The Statistics Notebook

Blocking accounts for known nuisance variation while preserving fair treatment comparisons. This page covers complete blocks, balanced incomplete blocks, and Latin squares. Numerical examples use illustrative practice data.

At a Glance¶

Complete blocks

Incomplete blocks

Each block contains every treatment. Treatment comparisons are made within blocks.

Randomized Complete Block Designs¶

Use a randomized complete block design (RCBD) when a known nuisance variable creates meaningful groups of similar units, and every treatment fits within every block.

Examples: compare production settings within each batch, or compare several conditions within each participant. Randomize within blocks, not across the whole experiment after forming them.

For $m$ treatments and $b$ blocks, with one observation per treatment-block cell,

Y_{ij}=\mu+\tau_i+\beta_j+\varepsilon_{ij},\qquad N=mb,

(1)

\sum_i\tau_i=0,\quad\sum_j\beta_j=0,\qquad \varepsilon_{ij}\overset{\mathrm{iid}}{\sim}N(0,\sigma^2).

(2)

Parameter	Meaning	Estimate in this balanced design
$\mu$	Overall mean	$\bar Y_{\cdot\cdot}$
$\tau_i$	Treatment deviation	$\bar Y_{i\cdot}-\bar Y_{\cdot\cdot}$
$\beta_j$	Block deviation	$\bar Y_{\cdot j}-\bar Y_{\cdot\cdot}$

\widehat Y_{ij}=\bar Y_{i\cdot}+\bar Y_{\cdot j}-\bar Y_{\cdot\cdot},\qquad e_{ij}=Y_{ij}-\widehat Y_{ij}.

(3)

This is an additive model. With only one observation per cell, treatment-by-block interaction cannot be separated from random error. The simple analysis assumes interaction is negligible. Repeated measurements can also require a covariance model beyond these independent errors.

Block ANOVA¶

\begin{aligned} SS_{\mathrm{Tr}}&=b\sum_i(\bar Y_{i\cdot}-\bar Y_{\cdot\cdot})^2,\\ SS_B&=m\sum_j(\bar Y_{\cdot j}-\bar Y_{\cdot\cdot})^2,\\ SS_E&=\sum_{i,j}(Y_{ij}-\bar Y_{i\cdot}-\bar Y_{\cdot j}+\bar Y_{\cdot\cdot})^2. \end{aligned}

(4)

SS_T=SS_{\mathrm{Tr}}+SS_B+SS_E.

(5)

Source	df	MS	Test
Treatment	$m-1$	$SS_{\mathrm{Tr}}/(m-1)$	$MS_{\mathrm{Tr}}/MSE$
Block	$b-1$	$SS_B/(b-1)$	$MS_B/MSE$
Error	$(m-1)(b-1)$	$SS_E/[(m-1)(b-1)]$
Total	$mb-1$

For $H_0:\tau_1=\cdots=\tau_m=0$ ,

F_{\mathrm{Tr}}\sim F_{m-1,(m-1)(b-1)}.

(6)

A fixed-block test uses $F_{b-1,(m-1)(b-1)}$ under $H_0:\beta_1=\cdots=\beta_b=0$ . The scientific focus is usually the treatment comparison. Blocking should be planned from knowledge of the experiment, not selected solely through a block $p$ -value.

A treatment contrast uses

\operatorname{SE}(\widehat C)=\sqrt{\frac{MSE}{b}\sum_i c_i^2},

(7)

with the block-model residual df. For two treatments, the treatment $F$ test equals the squared paired $t$ statistic.

Worked Blocking Example: Separating Treatment from Batch

Treatment	Batch 1	Batch 2	Batch 3	Batch 4	Mean
A	8	12	9	14	10.75
B	10	13	13	16	13.00
C	13	16	14	18	15.25

The grand mean is 13. For A in Batch 1,

\widehat y_{A1}=10.75+10.3333-13=8.0833,\qquad e_{A1}=-0.0833.

(8)

Source	SS	df	MS	$F$
Treatment	40.5000	2	20.2500	42.8824
Batch	52.6667	3	17.5556	37.1765
Error	2.8333	6	0.4722
Total	96.0000	11

The treatment-test $p$ -value is approximately 0.000280.

Ignoring batches puts their variation into the one-way residual:

MSE_{\mathrm{unblocked}}=\frac{52.6667+2.8333}{9}=6.1667.

(9)

Blocking reduces the error estimate to 0.4722 here. It also spends three residual degrees of freedom estimating batch effects. A block variable with little explanatory value need not improve precision.

The mean-difference estimates and simple orthogonal SS formulas above rely on the complete balanced layout. Missing cells or unequal replication require fitting the appropriate linear model; marginal mean subtraction is not generally enough.

Observed and batch-adjusted responses for three treatments across four batches. Adjustment reduces within-treatment spread while preserving the treatment means.

The right panel subtracts each estimated batch effect: $Y_{ij}-\widehat\beta_j$ . The treatment means stay the same in this balanced design, while the spread around each treatment mean shrinks. These adjusted responses illustrate the model; uncertainty still uses the block-model MSE and residual degrees of freedom.

Balanced Incomplete Block Designs¶

When each block can hold only $k<v$ treatments, a balanced incomplete block design (BIBD) spreads comparisons evenly.

Symbol	Meaning
$v$	Number of treatments.
$b$	Number of blocks.
$k$	Treatments per block, each occurring once in that block.
$r$	Blocks containing each treatment.
$\lambda$	Blocks containing each pair of distinct treatments.

Counting all assignments and treatment co-occurrences gives

bk=vr,\qquad r(k-1)=\lambda(v-1).

(10)

These are necessary conditions, not a guarantee that a design exists for arbitrary parameters. Under the usual additive model, balance makes treatment pair comparisons equally precise.

An incomplete block is incomplete by design, not a convenient description for uncontrolled missing observations. Allocate treatments and run order randomly within the design’s restrictions.

Latin Squares: Two Nuisance Factors¶

A $p\times p$ Latin square assigns $p$ treatments across two nuisance factors, each with $p$ levels. Every treatment appears once in each row and once in each column.

	Column 1	Column 2	Column 3	Column 4
Row 1	A	B	C	D
Row 2	B	C	D	A
Row 3	C	D	A	B
Row 4	D	A	B	C

Rows might be days and columns time windows. A letter tells us which treatment to apply at that row-column combination. There are $p^2$ runs, rather than all $p^3$ row-column-treatment combinations.

Randomly permute row assignments, column assignments, and treatment labels within an appropriate Latin-square randomization scheme. The displayed square is a layout template, not a substitute for randomization.

Let $k(i,j)$ identify the treatment in cell $(i,j)$ :

Y_{ij}=\mu+\rho_i+\gamma_j+\tau_{k(i,j)}+\varepsilon_{ij}.

(13)

Under the balanced additive model, row, column, and treatment effects have sum-to-zero constraints. With $\bar Y_k^{(\mathrm{Tr})}$ the mean of the $p$ observations receiving treatment $k$ ,

\begin{aligned} SS_R&=p\sum_i(\bar Y_{i\cdot}-\bar Y_{\cdot\cdot})^2,\\ SS_C&=p\sum_j(\bar Y_{\cdot j}-\bar Y_{\cdot\cdot})^2,\\ SS_{\mathrm{Tr}}&=p\sum_k(\bar Y_k^{(\mathrm{Tr})}-\bar Y_{\cdot\cdot})^2,\\ SS_E&=SS_T-SS_R-SS_C-SS_{\mathrm{Tr}}. \end{aligned}

(14)

Source	df	Treatment test
Rows	$p-1$
Columns	$p-1$
Treatment	$p-1$	$F=[SS_{\mathrm{Tr}}/(p-1)]/MSE$
Error	$(p-1)(p-2)$	$MSE=SS_E/[(p-1)(p-2)]$
Total	$p^2-1$

Under the treatment null, $F\sim F_{p-1,(p-1)(p-2)}$ under independent normal common-variance errors and additivity. For the $4\times4$ square, error df is 6. A $2\times2$ square has zero error df and cannot supply this error estimate.

The basic design cannot separately estimate general row-column, row-treatment, and column-treatment interactions. Strong interactions can invalidate the additive analysis.

Choosing the Analysis¶

Design or question	Analysis	Essential check
Two independent quantitative groups	Pooled or Welch $t$	Independent units; appropriate variance model.
Two matched conditions	Paired $t$ or within-pair randomization	Genuine pairing and independent pairs.
Several independent means	One-way ANOVA	Common variance and residual assumptions.
Specific mean combination	Contrast in the fitted model	Coefficients, SE, residual df, and testing family.
All pairwise mean differences	Tukey or another planned family procedure	Simultaneous rather than individual inference.
One nuisance factor; every treatment per block	RCBD	Randomization within blocks and additivity.
One nuisance factor; limited block capacity	Incomplete-block model	Connectivity, balance, and block-adjusted comparisons.
Two suitable nuisance factors	Latin square	One occurrence per row/column and negligible interactions.